Adjusting Soil Temperatures with a Physics-Informed Deep Learning Model for a High-Resolution Numerical Weather Prediction System

Soil temperature (ST) plays an important role in the surface heat energy balance, and an accurate description of soil temperatures is critical for numerical weather prediction; however, it is difficult to consistently measure soil temperatures. We developed a U-Net-based deep learning model to derive soil temperatures (designated as ST-U-Net) primarily based on 2 m air temperature (T2) forecasts. The model, the domain of which covers the Mt. Lushan region, was trained and tested by utilizing the high-resolution forecast archive of an operational weather research and forecasting four-dimensional data assimilation (WRF-FDDA) system. The results showed that ST-U-Net can accurately estimate soil temperatures based on T2 inputs, achieving a mean absolute error (MAE) of less than 0.8 K on the testing set of 5055 samples. The performance of ST-U-Net varied diurnally, with smaller errors at night and slightly larger errors in the daytime. Incorporating additional inputs such as land uses, terrain height, radiation flux, surface heat flux, and coded time further reduced the MAE for ST by 26.7%. By developing a boundary-layer physics-guided training strategy, the error was further reduced by 8.8%.